In this paper, we propose Cold-CFGAN, a collaborative filtering using two Conditional Generative Adversarial Networks (CGANs). In Cold-CFGAN, one CGAN is used for data augmentation of cold-start users, and the other CGAN is used to recommend items using user condition vectors. ColdCFGAN research uses an additional GAN model to generate data for cold-start users to resolve the cold start problem that occurs when implementing CGAN-based collaborative filtering and to further improve the accuracy of the model. To this end, we first identified the performance degradation problem
of cold-start users through a series of preliminary experiments using an existing conditional GAN-based collaborative filtering (CFGAN). Then, we used the user profile and item purchase data to express the number of purchased items per user in the form of a percentile, and identified cold-start users with few purchase items. Using the profile of the identified cold-start user data, we found the data of the Item-Rich user with the most similar profile to the cold-start user based on the cosine similarity, and using the data of the Item-Rich user, we applied partial masking method to create augmented cold-start users. Then we train user augmentation GAN to generate fake Item-Rich user using the augmented cold-start user and corresponding ItemRich user in real data. We use trained generator to generate Item-Rich user corresponding to cold-start user in real dataset. Then, we applied the generated Item-Rich user data to train the conditional GAN-based collaborative filtering and after training,
we performed experiment. Through the experiment, we found improved performance for cold start users compared to the traditional approach, and also improved overall performance.